The circular dependency hiding in codon references and GC targets

The question that exposed the loop

A codon-reference update raised a seemingly small question: should the existing GC target remain unchanged? The answer could not be separated from the reference itself.

The earlier target had not been anchored independently and then applied to the optimizer. It had been calibrated in part from outputs produced by the earlier reference table. The loop was therefore:

1. choose a codon table;
2. generate sequences with that table;
3. observe the output distribution; and
4. use that distribution to justify the range that later judges the outputs.

Changing only the reference caused the new outputs to be rejected by the old range. Changing only the range caused the old reference path to miss the new target. Either half-change made the other half look wrong.

The decision

The reference transition and the target-range revision had to be treated as one contract change. Historical results were not erased. They remained attached to the contract under which they were produced, while forward results used the new reference and target definition.

This led to a durable rule: a metric threshold should record its independent authority, not merely agree with the outputs of the system it evaluates.

Why this matters beyond GC

The same circularity can hide in other metrics. A CAI threshold, risk score, or ranking cutoff can appear objective while quietly inheriting assumptions from the generator. Separating generation authority from evaluation authority makes those dependencies visible and makes historical replay possible.

The native-composition range is also not the only legitimate design strategy. Explicit high-GC or other experimental profiles can remain available when labeled as non-default objectives. The important point is not that one range is universally correct; it is that the reason for selecting it is independently traceable.